Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Keywords: Two-dimensional phase unwrapping (2-D PU) has a strong influence on the accuracy of interferometric synthetic
Two-dimensional phase unwrapping (2-D PU) aperture radar (InSAR) data processing results. Phase gradient estimation (PGE) is one of the key steps in the
LinkNet with a pretrained encoder and dilated processing of 2-D PU. Moreover, the accuracy of the PGE will directly affect the accuracy of the final PU result. The
convolution (D-LinkNet)
phase continuity assumption is an important prerequisite for the PGE of traditional 2-D PU methods. However,
Interferometric synthetic aperture radar
the accuracy of PGE is not ideal in areas with high-noise and large-gradient changes. To address this issue, in
(InSAR)
Phase gradient estimation (PGE) this article, we propose a 2-D PU method of an unscented Kalman filter (UKF) using a refined LinkNet with a
pretrained encoder and dilated convolution (D-LinkNet). To the best of our knowledge, this article is the first time
to combine deep-learning and UKF for 2-D PU. First of all, this article analyzes the distribution characteristics
of different terrains. To ensure the accuracy of the training model, we use shuttle radar topography missions
(SRTMs) with different terrains to obtain simulated learning training data. Then, the refined d-LinkNet method
is used to accurately estimate the gradient ambiguity numbers and is combined with the small window median
filter to obtain the vertical and horizontal gradients. Finally, the UKF model is used for 2-D PU. Experiments are
conducted with simulated and TanDEM-X InSAR datasets. In addition, compared with the existing PGE methods
and 2-D PU methods, the experimental results show that the proposed method can obtain more accurate results
than the existing methods.
∗
Corresponding authors at: Land Satellite Remote Sensing Application Center, Ministry of Natural Resources of P.R.China, Haidian District, Beijing 100048, China.
E-mail addresses: wanggh@lasac.cn (G. Wang), wanggeng@satimage.cn (G. Wang).
https://doi.org/10.1016/j.optlaseng.2022.106948
Received 28 October 2021; Received in revised form 13 December 2021; Accepted 3 January 2022
0143-8166/© 2022 Elsevier Ltd. All rights reserved.
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
scholars have proposed to estimate the phase gradients with the maxi- and the development trends and research problems of deep-learning PU
mum likelihood (ML) [18] and the amended matrix pencil model [19]. are analyzed. The deep-learning method has been effectively applied in
These methods mainly estimate the terrain changes based on local fre- 2-D PU methods, which broadens the research direction of PU, and pro-
quency changes to obtain the phase gradients. These methods can obtain vides a good preliminary foundation for the research of deep-learning
high-precision phase gradient values from noisy interferometric phases. in the field of InSAR technology. However, the noise robustness of the
However, significant gradient phase loss occurs in high-noise areas with deep-learning PU method and the accuracy of the unwrapped results in
dense fringes [20]. PU in areas with large-gradient changes is currently a large-gradient change areas still need to be further studied.
popular research topic [21–24]. In response to this, multi-baseline (MB) From the foregoing, it can be seen that the PGE has an extremely
PU has received extensive attention. With the combination of the in- important impact on the 2-D PU accuracy and can even cause 2-D PU
terferometric phases of different baseline lengths, the range of phase failure, affecting the accuracy of the final obtained results. Although ex-
gradient ambiguity numbers is expanded to eliminate the limitation of isting methods have been used to carry out related research on PGE, PGE
the phase continuity assumption to obtain accurate phase gradients. Yu is still the bottleneck problem of 2-D PU in areas with high-noise and
and Lan used the Chinese remainder theorem (CRT) method combined large-gradient changes. In response to the above problems, in this arti-
with interferometric phases of different baseline lengths to estimate the cle, for the first time, we combine deep-learning and the UKF method for
phase gradients [25]. This method addresses the low PGE accuracy of PU. Firstly, we analyze different terrains, select the appropriate shuttle
traditional PU methods in large-gradient regions. Gao et al. combined radar topography missions (SRTMs) based on the different terrains, and
the CRT gradient estimation method with the unscented Kalman filter obtain an InSAR learning dataset. Then, the deep-learning PGE method
(UKF) method for PU [26]. The UKF can further correct the PGE error is analyzed. A refined d-LinkNet is used to estimate the vertical and hor-
and improve the accuracy of the 2-D PU. izontal phase gradient ambiguity coefficients and then is combined with
Kalman filter PU is a kind of path following-based method, and this a small window median filter to obtain the vertical and horizontal phase
method has better noise robustness. Loffeld et al. introduced Kalman gradients. Finally, the refined d-LinkNet PGE method is combined with
filter to PU and analyzed the extended Kalman filter (EKF) PU method the UKF model to perform PU. The accuracy of the proposed refined
in detail. In addition, the corresponding state and observation equa- d-LinkNet PGE method is significantly higher than that of the existing
tions are established, which further improves the accuracy of this type PGE methods. It is worthwhile mentioning that the proposed refined
of method [27]. However, the EKF model will lose high-order informa- d-LinkNet-based UKF-PU method can obtain better results than the ex-
tion when processing linearization problems. Therefore, Xie et al. intro- isting single-baseline method in areas with large-gradient changes.
duced iterated unscented Kalman filter [28] and unscented information This article is organized as follows. In Section 2, the principles of the
filtering [29] to the PU problem. Experimental results show that these 2-D PU methods based on the phase continuity assumption are reviewed,
methods can obtain better results. Ambrosino et al. introduced EKF to and the basic idea of how to detect the phase gradient information us-
the multichannel InSAR height reconstruction problem. The method can ing different methods is introduced. Then, the design of the refined d-
detect and correctly handle sharp height discontinuities while regulariz- LinkNet and the selected simulation InSAR dataset analysis used in this
ing smooth areas. Experimental results show that this method can obtain article are presented in Section 3. The refined d-LinkNet PGE method
better results [30]. and UKF-PU method are described in Section 4. In Section 5, the per-
Deep-learning technology has been widely used in fields such as formance of the proposed method is tested on both simulated and real
speech recognition and image segmentation. With the in-depth integra- InSAR datasets. The conclusions are presented in Section 6.
tion of deep-learning technology in various industries. In recent years,
deep-learning technology has been applied in 2-D PU. Schwartzkopf
2. Analysis and review
et al. [31] proposed a feed-forward multilayer perceptron neural net-
work that detects discontinuities based on the computation of proba-
InSAR observes the earth through microwave signals. It obtains mas-
bilities over a local patch. Zhang et al. [32] proposed a rapid and ro-
ter and slaver images of the same area and then conjugates the master
bust 2-D PU via deep-learning. This method has better efficiency and
and slaver images to obtain the interferometric phase [39]. At this mo-
noise robustness. Unfortunately, this method still cannot obtain ideal
ment, the obtained interferometric phase is wrapped in (−𝜋, 𝜋) In the
results in areas with large-gradient changes. Wang et al. [33] proposed
InSAR data processing process, the absolute phase can be obtained from
a one-step robust deep-learning PU, which uses deep-learning to im-
the wrapped phase. This process is named PU [40]. In the PU process,
prove the noise robustness of the unwrapped phase results. Spoorthi
PGE is the key research content of different PU methods. In this section,
et al. [34] proposed a deep convolutional neural network for a 2-D PU
we review the basic principle of 2-D PU, especially the PGE methods of
named PhaseNet. Although this method can obtain the ideal PU effi-
the traditional PU methods based on the phase continuity assumption.
ciency, it still cannot obtain ideal results in areas with large-gradient
changes. Then, PhaseNet2.0 was proposed [35]. Although this method is
a significant improvement over the previous methods, there are still sig- 2.1. The review of 2-D PU
nificant PU errors in the areas with large-gradient changes. Jiang et al.
[36] proposed a deep convolutional neural network PU for fringe projec- According to the above, the relationship between the absolute phase
tion 3D imaging, which can still obtain ideal results in high-noise areas. and wrapped phase in 2-D PU is as follows [41]:
Although the abovementioned methods can obtain unwrapped results
𝜓(𝑠) = 𝜑(𝑠) + 2𝑘(𝑠)𝜋 (1)
from interferometric phases with noise, they cannot obtain ideal results
in areas with large-gradient changes. Moreover, none of the above meth- where 𝜓(𝑠) represents the absolute phase of pixel 𝑠, 𝜑(𝑠) represents the
ods estimate the phase gradients, and the unwrapped results lack the wrapped phase of pixel 𝑠, and 𝑘(𝑠) represents the phase ambiguity num-
corresponding physical meaning. In view of the advantages and disad- ber of pixel 𝑠. The absolute phase gradient, i.e., the neighboring pixel
vantages of the above methods, Zhou et al. [37] first used deep-learning difference, can be obtained by:
methods to estimate the phase gradient ambiguity number and then used
Δ𝜓(𝑠, 𝑠 − 1) = 𝜓(𝑠 − 1) − 𝜓(𝑠)
the L1 -norm to obtain the unwrapped results. The experimental results
show that this method can still obtain ideal unwrapped results in high- = 𝜑(𝑠 − 1) − 𝜑(𝑠) + 2Δ𝑘(𝑠, 𝑠 − 1)𝜋
noise and large-gradient change areas. Subsequently, they conducted a = Δ𝜑(𝑠, 𝑠 − 1) + 2Δ𝑘(𝑠, 𝑠 − 1)𝜋 (2)
review analysis of the deep-learning PU methods and summarized the
existing deep-learning PU methods [38]. In this paper, the advantages where Δ𝜓(𝑠, 𝑠 − 1) represents the absolute phase gradient from pixel 𝑠 to
and disadvantages of deep-learning PU methods are introduced in detail, pixel 𝑠 − 1, Δ𝜑(𝑠, 𝑠 − 1) represents the wrapped phase gradient of pixel 𝑠,
2
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
3
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
Fig. 2. Illustration of the refined d-LinkNet-based PGE architecture. Wrapped phase image is given as input, and vertical and horizontal phase gradient ambiguity
numbers at each pixel are outputs.
Table 1
Major parameters of the refined d-Linknet-based PGE.
4
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
Fig. 4. The four representative terrain learning data used to train the model in this article. (a)–(d) Reference SRTM DEM of Litang, Weinan, Wuhan and Tianjin
respectively. (e)–(h)Simulated interferogram of (a)–(d) with 0.8 coherence coefficient.
prevent image resolution reduction and the loss of feature map spatial decoder part uses the transposed convolutional layer for upsampling and
information during the downsampling process of the interferogram. Ac- restores the resolution of the feature map from 16 × 16 to 256 × 256 to
cording to the single layer dilated convolution feature extraction of the obtain the phase gradient information.
interferogram, the input wrapped phase feature map is 𝜑(𝑠), and the di-
lated convolution 𝑦(𝑠, 𝑓 , 𝑟) at the position 𝑠 of the 𝑓 -th feature map can 3.2. Generation and introduction of the training setting
be expressed by the following formula:
Learning model training is a key step of the deep-learning PU
𝐸
∑ method. Therefore, appropriate training data are extremely important
𝑦(𝑠, 𝑓 , 𝑟) = 𝜑(𝑠 + 𝑟 ⋅ 𝑒) ⋅ 𝜃(𝑓 , 𝑒) + 𝑏(𝑠) (7)
for the accuracy and stability of the deep-learning PU method. The exist-
𝑒=1
ing deep-learning 2-D PU method uses simulation data for terrain sim-
where𝑟 is the dilation rate (in this article, the values are 1, 2, and 4), ulation or uses only the SRTM data of one type of terrain for simula-
𝐸 is the size of the convolution kernel, 𝜃(𝑓 , 𝑒) is the 𝑒-th parameter of tion learning. However, as we all know that the terrain is very complex
the 𝑓 -th convolution kernel, and 𝑏(𝑠) is the bias term. Therefore, the in actual situations, and it is difficult to guarantee the accuracy of the
cascaded dilated convolution of the center dilation can be expressed as: training model for deep-learning training using simulated data or a type
[ ] of terrain SRTM alone. According to the height range of the DEM, the
⎧ 𝑊∑ ⋅𝐻 ∑𝐹 terrain can be divided into plateaus, mountains, hills and plains. The
⎪𝑐 𝑟 ( 𝑦 ) = 𝑦(𝑠, 𝑓 , 𝑟)
⎨ 𝑠=1 𝑓 =1 (8) corresponding phase gradients of interferometric fringes of different to-
⎪𝑐 ( 𝑦 ) = 𝑐 ( 𝑐 ) pographies and geomorphologies are also different. Therefore, in order
⎩ 𝑟+𝑛 𝑟 +𝑛 𝑟
to ensure the accuracy of the deep-learning PU method training model,
where 𝑐𝑟 (𝑦) is the output gradient information feature map of the expan- in this article, four representative terrain SRTMs are selected for learn-
sion coefficient 𝑟, 𝑊 is the width of the feature map, 𝐻 is the height ing data simulation. The data are shown in Fig. 4. Fig. 4(a) is the SRTM
of the feature map,𝐹 is the number of convolution kernels, and 𝑐𝑟+𝑛 (⋅) in the Litang area of Sichuan, which has a large range of high mountains,
represents iterative calculation; that is, 𝑟 + 𝑛 represents the second con- and the topography starts relatively large and is obviously plateau to-
volution input and is the previous convolution result, where dilation pography. Fig. 4(b) shows the SRTM data of Weinan, Shaanxi, which has
rate is 𝑟. Compared with the original d-LinkNet, in this article, the ex- obvious mountain characteristics. Fig. 4(c) is the SRTM in the Wuhan
pansion coefficients of the dilated convolution layer are 1, 2 and 4, and area, which is an obvious hilly area, and Fig. 4(d) is the Tianjin area,
the receptive fields corresponding to each layer are 3, 7 and 15. The which is an obvious plain area. Then, we obtain the absolute phase and
input image size is 256 × 256, and after the encoding part, the feature simulate the interferometric phase. In addition, hypergeometric noise
map size is 16 × 16. In this case, in the dilated convolution with expan- with a coherence coefficient of 0.8 is added to the interferometric phase
sion coefficients of 1, 2 and 4, each output feature point will see 15 × 15 to obtain the input data of the training model. According to Eqs. (4) and
feature points on the input feature map. This will cover the main part (6), the different terrain simulation data are processed to obtain the
of the input interferogram feature map and captures the global gradient phase ambiguity numbers of different terrains, and the phase ambiguity
feature information. numbers are used as the output data of the training model. In this article,
The decoder used by the refined d-LinkNet in this article is the same we use 13,008 interferograms for training the network and validating
as the original LinkNet, which can ensure computational efficiency. The the network. 80% of the interferograms used for training the network
5
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
and 20% of the interferograms used for validating the network. Finally,
our network is trained on an NVIDIATM Tesla v100 with 32 GB of GPU
memory. The statistical training time is 12632s.
4.1. PGE using refined d-LinkNet areas with large-gradient changes. The UKF 2-D PU basic equation is as
follows:
According to the previous article, when the refined d-LinkNet has
been trained, it can be employed to predict the phase gradient infor-
mation of the wrapped phase image. In this article, we simulate the ⎧𝜓(𝑠) = 𝜓(𝑠 − 1) + 𝛿(𝑠 − 1) + 𝜔(𝑠 − 1)
⎪ { } { }
SRTM data of different terrains and train the model (as shown Fig. 4). ⎨𝜙(𝑠) = sin(𝜓(𝑠)) + 𝑣1 (𝑠) = ℎ[𝜓(𝑠)] + 𝑣(𝑠) (10)
With the addition of different types of noise to the simulated data, the ⎪ cos(𝜓(𝑠)) 𝑣2 (𝑠)
⎩
refined d-LinkNet PGE method can recognize credible phase gradient
information from complicated and noisy interferograms. Therefore, ac-
cording to Eq. (2), we can obtain the vertical and horizontal gradients where 𝜓(𝑠 − 1) represents the absolute phase of the 𝑠 − 1th pixel, 𝛿(𝑠 − 1)
as follows: is the absolute phase gradient, 𝜔(𝑠 − 1) is the PGE error of the 𝑠 − 1th
{ pixel, ℎ[⋅] and 𝜙(𝑠) are the observation coefficient vector and observa-
Δ𝜓̄ 𝑉 (𝑠, 𝑠 − 1) = Δ𝜑𝑉 (𝑠, 𝑠 − 1) + 2Δ𝑘𝑉 (𝑠, 𝑠 − 1)𝜋
Δ𝜓̄ 𝐻 (𝑠 − 1, 𝑠) = Δ𝜑𝐻 (𝑠 − 1, 𝑠) + 2Δ𝑘𝐻 (𝑠 − 1, 𝑠)𝜋
(9) tion value, respectively, and 𝑣1 (𝑠) and 𝑣2 (𝑠) are the observation vari-
ances. Eq. (10) shows that the UKF 2-D PU takes the PGE error as the
where Δ𝜓̄ 𝑉 andΔ𝜓̄ 𝐻 are the PGEs of the vertical and horizontal direc- state equation variance. Therefore, it can further improve the accuracy
tions, respectively (using a 5 × 5 window median filter for slight fil- of refined d-LinkNet PGE. The 2-D UKF PU model can be divided into
tering), Δ𝜑𝑉 and Δ𝜑𝐻 are the vertical and horizontal wrapped phase three steps. The first step is the state equation prediction. The prediction
differences, respectively, and Δ𝑘𝑉 andΔ𝑘𝐻 are the vertical and hori- equation is:
zontal phase gradient ambiguity numbers from the refined d-LinkNet,
respectively.
⎧𝜒 − (𝑠) = 𝑓 [𝜉 (𝑠 − 1)]
⎪ 𝑗 𝑗
4.2. Introduction of UKF PU ⎪ ∑2
⎪𝜓⃗ − (𝑠) = 𝑏𝑚𝑗 𝜒𝑗 ( 𝑠 )
−
⎨ 𝑗=0
(11)
UKF 2-D PU not only has better noise robustness, but also can cor- ⎪ [ ][ ]𝑇
⎪𝑃 − ( 𝑠 ) = ∑ 𝑏 𝑐 𝜒 − ( 𝑠 ) − 𝜓
2
rect the phase gradients and improve the accuracy of the unwrapped ⃗ − (𝑠) 𝜒𝑗− (𝑠) − 𝜓
⃗ − (𝑠) + 𝑄(𝑠 − 1)
phase. UKF 2-D PU is a path following-based method. Different from the ⎪ 𝜓𝜓 𝑗=0
𝑗 𝑗
⎩
traditional 2-D PU method, UKF 2-D PU obtains the unwrapped phase
based on an efficient quality-guided strategy based on a heapsort. This
where 𝜉𝑗 (𝑠 − 1)is the sigma point of the state estimate at pixel 𝑠 − 1. In
strategy ensures that the PU path is always going from a high-quality √
this article, 𝜉0 (𝑠 − 1)=𝜓 ⃗ (𝑠 − 1), 𝜉1 (𝑠 − 1)=𝜓⃗ (𝑠 − 1)+{ (1 + 𝜆)𝑃𝜓𝜓 (𝑠 − 1)},
point to a low-quality point in an area. As we know, an area with large- √
gradient changes, that is, a discontinuous phase, has low-quality in the and 𝜉2 (𝑠 − 1)=𝜓 ⃗ (𝑠 − 1) − { (1 + 𝜆)𝑃𝜓𝜓 (𝑠 − 1)}𝜓 ⃗ (𝑠 − 1) and 𝑃𝜓𝜓 (𝑠 − 1)
quality map. Therefore, the UKF 2-D PU can obtain the unwrapped phase represent the state estimate and its corresponding state estimation co-
through the strategy of "circumnavigation and climbing" in areas with variance, respectively. 𝜒𝑗− (𝑠) represents the predicted values of the
large-gradient changes. As shown in Fig. 5, the orange point is the un- sigma points. 𝑏𝑚 𝑐 𝑚
𝑗 and 𝑏𝑗 refer to the weight factors, 𝑏0 = 𝜆∕(1 + 𝜆),
𝑐 𝑚 𝑐
𝑏0 = 𝜆∕(1 + 𝜆) + (1 − 𝜔 + 𝑙), 𝑏𝑗 = 𝑏𝑗 = 1∕2(1 + 𝜆), (𝑗 = 1, 2), 𝜆 is the pa-
2
wrapped point, and the gray point is the point to be unwrapped. The
red line is the phase discontinuity line. If the path of PU passes through rameter used to adjust the sigma points, and 𝜆 = 𝜔2 (1 + 𝜅) − 1. In this ar-
the red line, the unwrapped result produces a phase jump. Most of the ticle, 𝜔 = 0.01, 𝑙 = 2, and 𝜅 = 0. 𝜓 ⃗ − (𝑠)and𝑃𝜓𝜓
− (𝑠)are the predicted value
existing PU methods use the 2nd path for PU. In comparison, the UKF of the unwrapped phase and its corresponding prediction error covari-
PU method used in this article can use the 1st or 3rd path to obtain the ance matrix, respectively, and𝑄(𝑠 − 1) represents the variance in the
unwrapped phase of the gray point. Therefore, the UKF 2-D PU can ob- phase error of the 𝑠 − 1th pixel. According to Eqs. (10) and (11), we
tain higher precision results than the existing single-baseline 2-D PU in can obtain the predicted value of the observation value of the second
6
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
⎪
⎪𝜙− (𝑠) = ∑ 𝑏𝑚 𝜁 − (𝑠)
2
⎨ 𝑗=0
𝑗 𝑗 (12)
⎪ [ ] [ ]
⎪𝑃 − (𝑠) = ∑ 𝑏𝑐 𝜁 − (𝑠) − 𝜙− (𝑠) ⋅ 𝜁 − (𝑠) − 𝜙− (𝑠) 𝑇 + 𝑅(𝑠)
2
⎪ 𝜙𝜙 𝑗 𝑗 𝑗
⎩ 𝑗=0
where 𝜁𝑗− (𝑠) represents the predicted value. 𝜙− (𝑠) and 𝑃𝜙𝜙 − (𝑠) repre-
sent the predicted values of the unwrapped phase at the 𝑠th pixel
and its corresponding prediction error covariance matrix, respectively.
𝑅(𝑠)represents the observation variance. Assume that the unwrapped
phase estimate at a pixel and its corresponding estimate error variance
are 𝜓(𝑠) and 𝑃𝜓𝜓 (𝑠), respectively. Thus, we can obtain the following:
⎧ [ ] [ ]𝑇
− (𝑠) = ∑ 𝑏𝑐 𝜒 − (𝑠) − 𝜓 − (𝑠) ⋅ 𝜁 − (𝑠) − 𝜙− (𝑠)
2
⎪ 𝑃𝜓𝜙 𝑗 𝑗 𝑗
⎪ 𝑗=0
⎪ 𝓁(𝑠) = 𝑃 − (𝑠)∕𝑃 − (𝑠)
⎨ 𝜓𝜙 𝜙𝜙 (13)
⎪ 𝜓(𝑠) = 𝜓 ⃗ − (𝑠) + 𝓁(𝑠) ⋅ [𝜙(𝑠) − 𝜙− (𝑠)]𝑇 Fig. 6. (a) Reference DEM. (b) Simulated noise-free wrapped phase image. (c)
⎪ Simulated wrapped phase image with 0.7 coherence coefficient.
⎪𝑃𝜓𝜓 (𝑠) = 𝑃 − (𝑠) − 𝓁(𝑠) ⋅ 𝑃 − (𝑠) ⋅ 𝓁 (𝑠)𝑇
⎩ 𝜓𝜓 𝜓𝜙
where 𝑃𝜓𝜙− (𝑠) represents the covariance of the predicted value of the 𝑠th
pixel, 𝜙(𝑠) and 𝜙− (𝑠) are the actual and predicted values, respectively, [37] that PGNet can obtain the phase gradient ambiguity numbers from
𝓁(𝑠) denotes the gain matrix of the 𝑠th pixel, and 𝜓(𝑠) and 𝑃𝜓𝜓 (𝑠) rep- an interferometric phase with low coherence. Moreover, the experimen-
resent the state estimate of the interferometric pattern of the 𝑠th pixel tal results show that the phase gradients obtained by the PGNet method
and the corresponding variance in the estimation error, respectively. are more accurate than the phase continuity assumption. The PGNet
According to Eqs. (10)–(13), we can obtain the unwrapped phase. It method provides a good technical means for the PGE of 2-D PU. How-
worth to remind that the above equations are all introduced in one- ever, the accuracy of the phase gradient information estimated by the
dimensional algorithmic formulas. Therefore, with the change of 𝑠to 2- PGNet method still needs to be further improved. In view of this, this
D, the above equations are the 2-D PU model. The 2-D UKF PU has article uses a refined d-LinkNet to estimate the vertical and horizontal
been proposed, however, the 2-D UKF-PU combined with refined d- phase gradient ambiguity numbers of the interferometric phase. In this
LinkNet has never been mentioned. Moreover, the PU characteristics section, we use the simulation data of Fig. 6(c) to conduct experiments
of 2-D UKF-PU in large-gradient areas have not been reflected in rele- and to compare the accuracy of the PGE results of the phase continuity
vant studies. Therefore, this paper proposes a refined d-LinkNet-based assumption, PGNet and the proposed refined d-LinkNet-based method.
UKF-PU method, which that optimizes and improves the existing UKF The PGE results obtained by different methods are illustrated in
PU method and gives a detailed introduction to the characteristics of Fig. 7. Fig. 7(a) and (b) are the true vertical and horizontal phase gra-
its PU. dient ambiguity numbers of Fig. 6(b) obtained by the traditional phase
continuity assumption. Fig. 6(b) is the interferometric phase without
noise, so the PGE in Fig. 7(a) and (b) is the true phase gradient ambigu-
5. Performance analysis ity numbers. Fig. 7(c) and (d) are the vertical and horizontal phase gra-
dient ambiguity numbers of Fig. 6(c) obtained by the phase continuity
In this section, we conduct the performance evaluation and illustra- assumption. The PGE results obtained by the traditional phase continu-
tion of the proposed refined d-LinkNet-based UKF-PU method in com- ity assumption contain obvious noise. Fig. 7(e) and (f) are the vertical
parison with several representative 2-D PU methods through three ex- and horizontal phase gradient ambiguity numbers of Fig. 6(c) obtained
periments. The first experiment demonstrates the PGE robustness of the by PGNet. It can be seen from the figure that the PGNet method can
refined d-LinkNet-based UKF-PU on the simulated InSAR dataset com- obtain better PGE results than the phase continuity assumption. How-
pared with that of the existing PGE. The second experiment is also per- ever, there will still be some residual noise, especially between the two
formed on the simulated InSAR dataset, which validates the effective- lines. According to phase gradients, we can compute the residue dis-
ness of the proposed method using the phase gradients from the re- tribution. Fig. 7(c) and (d) obtained by phase continuity assumption
fined d-LinkNet regarding obtaining the final PU result. The third ex- contains 42,485 residuals, Fig. 7(e) and (f) obtained by PGNet con-
periment is performed on an actual InSAR dataset to compare refined tains 37,597 residuals. Although PGNet can obtain better results than
d-LinkNet-based UKF-PU with the MCF, SNAHPU and PGNet 2-D PU the phase continuity assumption, there will still be significant noise
methods. residues. Moreover, as the noise content increases, the residual noise
in the PGE result of the PGNet method will also increase. Fig. 7(g) and
5.1. Evaluation of the robustness of d-LinkNet PGE (h) are the vertical and horizontal phase gradient ambiguity numbers
of Fig. 6(c) obtained by refined d-LinkNet-based method. From the re-
In the first experiment, we test the robustness of the refined d- sults, we can see that the refined d-LinkNet-based method proposed in
LinkNet-based PGE method in detecting the phase gradient ambiguity this article can obtain better PGE results than the phase continuity as-
numbers on the simulated InSAR data. Fig. 6(a) shows the reference sumption and PGNet. Moreover, the refined d-LinkNet-based method
terrain height, which is also from the SRTM DEM in the Litang area of basically has no residual noise phenomenon. Compared with Fig. 7(a)
Sichuan, China. Fig. 6(b) shows a simulated noise-free wrapped phase and (b), it is found that the vertical and horizontal phase gradient am-
image. Fig. 6(c) shows a simulated wrapped phase image with a coher- biguity numbers estimated by the refined d-LinkNet-based method are
ence coefficient of 0.7. According to the existing research, PGE has an basically consistent with the true phase gradients. It can be seen that the
extremely important influence on 2-D PU. In [37], PGNet was the first proposed refined d-LinkNet-based method is an effective high-precision
application of deep-learning to the PGE problem. It can be seen from PGE method.
7
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
Table 2
Evaluation results of simulation dataset.
unwrapped phase of Fig. 8(d). We use MCF, SNAHPU, PGNet and the
refined d-LinkNet-based UKF-PU to wrapped the interferometric phases
of four different terrains, respectively. It is worth mentioning that in
the simulation data experiment, we find that the PU error caused by
the large-gradient change is weakened. Therefore, the simulation data
experiment can compare the noise robustness of different PU methods
in different terrain PU problems. To further analyze the accuracy of dif-
ferent PU methods, we carry out statistical analysis on the unwrapped
results of different PU methods. As shown in Table 2 and Fig. 9, for the
PU results of Fig. 8(a), the root means square errors (RMSEs) of the MCF
and SNAPHU methods are 0.4260 and 0.4262 rad, respectively, which
reflect poor accuracy. The RMSE of PGNet is 0.4226; although the accu-
racy is better than those of the other two methods, it can be seen from
Fig. 9(a) that the PU error distribution of PGNet is relatively discrete.
The RMSE of the proposed refined d-LinkNet-based UKF-PU method is
0.1389 rad. The accuracy of the PU result is the best, and its PU error dis-
tribution map is more concentrated, and the error distribution is smaller
than those of the other three methods. Table 2 shows that the RMSEs of
the PU results of MCF and SNAPHU in Fig. 8(b) are 0.5405 and 0.5403,
respectively. The RMSE of PGNet is 0.6511. The RMSE of the proposed
refined d-LinkNet-based UKF-PU method is 0.3847 rad. In mountain-
ous areas, the accuracy of the PU results of the refined d-LinkNet-based
UKF-PU method proposed in this paper is still better than those of the
other three methods. From the error distribution graph in Fig. 9(b), we
can further see that the PU result of the refined d-LinkNet-based UKF-
PU is the best. The accuracy and error distribution of the PU results
of Fig. 8(c) are very similar to the results of Fig. 8(a), and the refined
Fig. 7. (a) - (b)Vertical and horizontal phase gradients image of Fig. 6(b) ob- d-LinkNet-based UKF-PU can still obtain ideal PU results. In plain ar-
tained by the phase continuity assumption. (c) - (d) Vertical and horizontal phase eas, the difficulty of PU is greatly reduced; however, the accuracy of the
gradient ambiguity numbers image of Fig. 6(c) obtained by the phase continuity PU result is still affected due to the influence of noise. For example, in
assumption. (e) - (f) Vertical and horizontal phase gradient ambiguity numbers Fig. 8(d), the RMSEs of the MCF and SNAPHU methods are 0.4224 and
image of Fig. 6(c) obtained by PGNet. (g) - (h)Vertical and horizontal phase 0.4226 rad, respectively. Although the RMSE of PGNet is 0.4186 rad,
gradient ambiguity numbers image of Fig. 6(c) obtained by refined d-LinkNet. which is a higher accuracy than those of the other two methods, it can
be seen from Fig. 9(d) that its error distribution is still relatively large.
The RMSE of the proposed refined d-LinkNet-based UKF-PU method is
only 0.1231 rad, which is still the best accuracy of the PU methods. The
5.2. Simulation-Based 2-D PU performance analysis
effectiveness of the proposed refined d-LinkNet-based UKF-PU method
is proven through a simulation data experiment on four different ter-
In the second experiment, we will examine the PU effectiveness of
rains.
the refined d-LinkNet-based UKF-PU by using four different terrains in
Fig. 4 as the test simulation data. We select four pieces of simulated
data from four different terrains in Fig. 4 for data experiments. The 5.3. Validation of PU performance on real case
performance of the refined d-LinkNet-based UKF-PU method is tested
against the other 2-D PU methods on a simulated InSAR dataset, which To further verify the performance of the refined d-LinkNet-based
is shown in Fig. 8. Fig. 8(a) shows the interferometric phase intercepted UKF-PU method proposed in this article, a TanDEM-X InSAR dataset is
in Fig.4(e). Fig. 8(a) shows that the interferometric phase we intercepted used for verification. Fig. 10(a) is a Google Earth image of the observed
has certain dense fringe areas, which is an obvious topographic feature area. Fig. 10(b) shows the TanDEM-X interferogram, and Fig. 10(c) is the
of plateau and mountain areas. Fig. 8(b) shows the interferometric phase external SRTM DEM. Fig. 10(d) is the reference phase from Fig. 10(c).
intercepted in Fig. 4(f). It can be seen that the fringes in Fig. 8(b) are The Google Earth image shows many mountains in the study area, and
density, which is consistent with the fringe characteristics in mountain- hence, there are large terrain gradients. Fig. 10(e)–(h) presents the PU
ous areas. This will also increase the difficulty of PU. Fig. 8(c) shows results of the MCF, SNAPHU, PGNet, and proposed refined d-LinkNet-
the interferometric phase intercepted in Fig. 4(g). Fig. 8(d) shows the based UKF-PU methods, respectively. Fig. 10(i)–(l) shows the PU er-
interferometric phase intercepted in Fig. 4(h). Fig. 8(e) is the reference rors between Fig. 10(d) and (e)–(h). It can be seen from the figure
unwrapped phase of Fig. 8(a). Fig. 8(f) is the reference unwrapped phase that due to the influence of large-gradient changes, MCF, SNAPHU and
of Fig. 8(b). It is a typical interference phase of hilly terrain. Fig. 8(g) PGNet all produce obvious PU errors, which seriously affects the accu-
is the reference unwrapped phase of Fig. 8(c). Fig. 8(h) is the reference racy of the PU results. The PU result of the proposed d-LinkNet-based
8
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
Fig. 8. PU performance comparison between d-LinkNet-based UKF-PU and representative PU methods on the simulated data set. (a) Interferometric phase intercepted
in Fig. 4(e). (b) Interferometric phase intercepted in Fig. 4(f). (c) Interferometric phase intercepted in Fig. 4(g). (d) Interferometric phase intercepted in Fig. 4(h). (e)
Reference unwrapped phase of Fig. 8(a). (f) Reference unwrapped phase of Fig. 8(b). (g) Reference unwrapped phase of Fig. 8(c). (h) Reference unwrapped phase of
Fig. 8(d).
Fig. 9. Error distribution diagram of the simulated data set results. The horizontal axis is estimate error, and the vertical axis is number of pixels. (a) is the error
distribution diagram of the unwrapped result of Fig. 8(a). (b) is the error distribution diagram of the unwrapped result of Fig. 8(b). (c) is the error distribution
diagram of the unwrapped result of Fig. 8(c). (d) is the error distribution diagram of the unwrapped result of Fig. 8(d).
Table 3 Table 4
Major parameters of the TanDEM-X dataset. Evaluation results of TanDEM-X dataset.
Orbit Altitude Incidence Angle Wavelength Normal Baseline PU method Figure RMSE (rad)
9
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
Fig. 10. PU performance comparison between d-LinkNet-based UKF-PU and representative PU methods on the TanDEM-X InSAR dataset. (a) Google Earth image of
the observed area. (b) TanDEM-X interferogram. (c) The external DEM. (d) Reference PU result of (b) obtained from the external DEM. (e)–(h) are the PU results of
(b) obtained by MCF, SNAPHU, PGNet and d-LinkNet-based UKF-PU, respectively. (i)–(l) are differences between (d) and (e)–(h).
ularly important to obtain high-precision phase gradients in high-noise releases from patients whose names or photographs are submitted as
areas. Large-gradient changes have always been a key issue of 2-D PU, part of the work.
especially for single-baseline PU. The proposed method uses a refined
d-LinkNet to estimate the phase gradient ambiguity coefficient and ob- Declaration of Competing Interest
tains high-precision phase gradient information from the interferometric
phase with noise. To ensure the accuracy of the refined d-LinkNet train- None.
ing model, four different terrain SRTMs are selected for training data
simulation in the article. Then, the phase gradient ambiguity coefficient
CRediT authorship contribution statement
estimated by the refined d-LinkNet and the median filter of the small
window are combined to obtain the vertical and horizontal gradients.
Yandong Gao: Conceptualization, Methodology, Software, Valida-
Finally, the UKF based on heapsort unwraps the interferometric phase.
tion, Writing – original draft, Writing – review & editing, Funding ac-
The performance of the refined d-LinkNet-based UKF-PU method is eval-
quisition. Guanghui Wang: Conceptualization, Validation, Writing – re-
uated using simulations and real data experiments, and the results show
view & editing, Project administration. Geng Wang: Validation, Formal
that the method proposed in this paper has better noise robustness than
analysis, Project administration. Shubi Zhang: Investigation, Supervi-
the existing PU methods and can obtain better unwrapped results than
sion. Shijin Li: Resources, Supervision.
other methods in areas with large-gradient changes.
Each author certifies that he or she has participated sufficiently in This work was funded by the National Natural Science Founda-
the preparation of the present work to take public responsibility for it. tion of China (No: 42001409), China Postdoctoral Science Founda-
Each has reviewed the final version of the work, believes it is a valid tion (No: 2020M681770), National Key R&D Program of China (No:
work, and approves it for publication. 2017YFE0119600), and Development Fund of the Key Laboratory of
Land Satellite Remote Sensing Application Center, Ministry of Natural
Disclaimer Resources of P.R. China (No: KLSMNR-202103). The authors also thank
the anonymous reviews for their constructive comments and sugges-
Each author warrants that this work contains no libelous or unlawful tions.
statements and does not infringe on the rights of others. If excerpts (text,
figures, tables, or illustrations) from copyrighted works are included, a References
written release will be secured by the authors prior to submission, and
[1] Noviello C, et al. Monitoring buildings at landslide risk with SAR: a methodology
credit to the original publication will be properly acknowledged. Each based on the use of multipass interferometric data. IEEE Geosci Remote Sens Mag
author warrants that he or she has obtained, prior to submission, written (replaces Newsletter) 2020;8(1):91–119.
10
Y. Gao, G. Wang, G. Wang et al. Optics and Lasers in Engineering 152 (2022) 106948
[2] Rosen PA, et al. Synthetic aperture radar interferometry. Proc IEEE [25] Yu H, Lan Y. Robust two-dimensional phase unwrapping for multibaseline SAR in-
2000;88(3):333–82. terferograms: a two-stage programming approach. IEEE Trans Geosci Remote Sens
[3] Ghiglia DC, Mastin GA, Romero LA. Cellular automata method for phase unwrap- 2016;54(9):5217–25.
ping. J Opt Soc Am 1987;4(1):267–80. [26] Gao Y, Zhang S, Li T, Chen Q, Zhang X, Li S. Refined two-stage programming ap-
[4] Yu H, Li Z, Bao Z. A cluster-analysis-based efficient multibaseline phase-unwrapping proach of phase unwrapping for multi-baseline SAR interferograms using the un-
algorithm. IEEE Trans Geosci Remote Sens 2011;4(1):478–87. scented Kalman filter. Remote Sens 2019;11:199.
[5] Ghiglia DC, Pritt MD. Two-dimensional phase unwrapping: theory, algorithms, and [27] Loffeld O, Nies H, Knedlik S, Yu W. Phase unwrapping for SAR interferome-
software. New York: Wiley-Interscience; 1998. try-a data fusion approach by Kalman filtering. IEEE Trans Geosci Remote Sens
[6] Yu H, Lan Y, Yuan Z, Xu J, Lee H. Phase unwrapping in InSAR: a review. IEEE Geosci 2008;46(1):47–58.
Remote Sens Mag 2019;7(1):40–58. [28] Xie X. Iterated unscented Kalman filter for phase unwrapping of interferometric
[7] Dai Y, Ng AHM, Wang H, Li L, Ge L, Tao T. Modeling assisted InSAR phase-un- fringes. Opt Express 2016;24(17) pp. 18,872-97.
wrapping method for mapping mine subsidence. IEEE Geosci Remote Sens Lett [29] Xie X, Dai G. Unscented information filtering phase unwrapping algorithm for inter-
2021;18(6):1059–63. ferometric fringe patterns. Appl Opt 2017;56(34):9423–34.
[8] Gao Y, Tang X, Li T, Lu J, Li S, Chen Q, Zhang X. A phase slicing 2-D phase unwrap- [30] Ambrosino R, Baselice F, Ferraioli G, Schirinzi G. Extended Kalman filter for
ping method using the L1 -Norm. IEEE Geosci Remote Sens Lett 2020 to be published. multichannel InSAR Height Reconstruction. IEEE Trans Geosci Remote Sens
doi:10.1109/LGRS.2020.3025939. 2017;55(10):5854–63.
[9] Goldstein RM, Zebker HA, LWerner C. Statellite radar nterferometry: two-dimen- [31] Schwartzkopf W, Milner TE, Ghosh J, Evans BL, Bovik AC. Two-dimensional phase
sional phase unwrapping. Radio Sci 1988;23(4):713–20. unwrapping using neural networks. In: Proceedings of the 4th IEEE Southwest sym-
[10] Zhong H, Tang J, Zhang S, Zhang X. A quality-guided and local minimum discontinu- posium on image analysis and interpretation. Austin, TX, USA; 2000. p. 247–50.
ity-based phase unwrapping algorithm for insar/insas interferograms. IEEE Geosci [32] Zhang T, et al. Rapid and robust two-dimensional phase unwrapping via deep learn-
Remote Sens Lett 2014;11(1):215–19. ing. Opt Express 2019;27(16) pp.23,173–23,185, July.
[11] Xu J, An D, Huang X, Yi P. An efficient minimum-discontinuity phase-unwrapping [33] Wang K, Li Y, Kemao Q, Di J, Zhao J. One-step robust deep learning phase unwrap-
method. IEEE Geosci Remote Sens Lett 2016;13(5):666–70. ping. Opt Express 2019;27(10) pp. 15,100–15,115.
[12] Dai Z, Zha X. An accurate phase unwrapping algorithm based on reliablity sorting [34] Spoorthi GE, Gorthi S, Gorthi RKSS. PhaseNet: a deep convolutional neural network
and residue mask. IEEE Geosci Remote Sens Lett 2012;9(2):219–23. for two-dimensional phase unwrapping. IEEE Signal Process Lett 2019;26(1):54–8.
[13] Costantini M. A novel phase unwrapping method based on network programming. [35] Spoorthi GE, Gorthi RKSS, Gorthi S. PhaseNet 2.0: phase unwrapping of noisy data
IEEE Trans Geosci Remote Sens 1998;36(3):813–21. based on deep learning approach. IEEE Trans Image Process 2020;29:4862–72.
[14] Lan Y, Yu H, Xing M. Refined two-stage programming-based multi-baseline phase [36] Liang J, Zhang J, Shao J, Song B, Yao B, Liang R. Deep convolutional neural network
unwrapping approach using local plane model. Remote Sens 2019;11:491. phase unwrapping for fringe projection 3D imaging. Sens (Basel) Jul. 2020;20.
[15] Chen CW, Zebker HA. Two-dimensional phase unwrapping with use of statistical [37] Zhou L, Yu H, Lan Y. Deep convolutional neural network-based robust phase gradient
models for cost functions in nonlinear optimization. J Opt Soc Am A Opt Image Sci estimation for two-dimensional phase unwrapping using SAR interferograms. IEEE
2001;18(2):338–51. Trans Geosci Remote Sens 2020;58(7):4653–65.
[16] Xie X, Zeng Q. Efficient and robust phase unwrapping algorithm based on unscented [38] Zhou L, Yu H, Lan Y, Xing M. Artificial intelligence in interferometric synthetic
Kalman filter, the strategy of quantizing paths-guided map, and pixel classification aperture radar phase unwrapping. IEEE Geosci Remote Sens Mag 2021;9(2):10–28.
strategy. Appl Opt 2015;54(31):9294–307. [39] Moreira A, Prats-Iraola P, Younis M, Krieger G, Hajnsek I, Papathanassiou KP. A
[17] Gao Y, Zhang S, Li T, Chen Q, Li S, Meng P. Adaptive unscented kalman filter phase tutorial on synthetic aperture radar. IEEE Geosci Remote Sens Mag 2013;1(1):6–43.
unwrapping method and its application on gaofen-3 interferometric SAR data. Sens [40] Yu H, Xing M, Bao Z. A fast phase unwrapping method for large-scale interferograms.
(Basel) 2018;18. IEEE Trans Geosci Remote Sens 2013;51(7):4240–8.
[18] Liu W, Bian Z, Liu Z, Zhang Q. Evaluation of a cubature Kalman filtering-based phase [41] Yu H, Lan Y, Xu J, An D, Lee H. Large-scale L0-norm and L1-norm two-dimensional
unwrapping method for differential interferograms with high noise in coal mining phase unwrapping. IEEE Trans Geosci Remote Sens 2017;55(8):4712–28.
areas. Sens (Basel) 2015;15:16336–57. [42] Yu H, Zhou Y, Ivey SS, Lan Y. Large-scale multibaseline phase unwrapping: interfer-
[19] Xie X, Li Y. Enhanced phase unwrapping algorithm based on unscented Kalman ogram segmentation based on multibaseline envelope-sparsity theorem. IEEE Trans
filter, enhanced phase gradient estimator, and path following strategy. Appl Opt Geosci Remote Sens 2019;57(11):9308–22.
2014;53(18):4049–60. [43] Liu H, Xing M, Bao Z. A cluster-analysis-based noise-robust phase-unwrapping
[20] Gao Y, Zhang S, Li T, Guo L, Chen Q, Li S. A novel two-step noise reduction approach algorithm for multibaseline interferograms. IEEE Trans Geosci Remote Sens
for interferometric phase images. Opt Lasers Eng 2019;121:1–10. 2015;53(1):494–504.
[21] Fornaro G, Guarnieri AM, Pauciullo A, De-Zan F. Maximum likelihood multi-baseline [44] Yu H, Lee H. A convex hull algorithm based fast large-scale two-dimensional phase
SAR interferometry. IEE Proc Radar Sonar Navigat Jun. 2006;153(3):279–88. unwrapping method. In: Proceedings of the IEEE international geoscience and re-
[22] Ferraiuolo G, Pascazio V, Schirinzi G. Maximum a posteriori estimation of height mote sensing symposium. Fort Worth, Texas, USA; 2017. p. 3824–7.
profiles in InSAR imaging. IEEE Geosci Remote Sens Lett 2004;1(2):66–70. [45] Zhou L, Zhang C, Wu M. D-LinkNet: LinkNet with pretrained encoder and dilated
[23] Xu W, Chien Chang E, Keong Kwoh L, Lim H, Cheng W, Heng A. Phase-unwrap- convolution for high resolution satellite imagery road extraction. In: Proceedings of
ping of SAR interferogram with multifrequency or multi-baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) workshops;
the IEEE international geoscience and remote sensing symposium (IGARSS); 1994. 2018.
p. 730–2. [46] Chaurasia A, Culurciello E. LinkNet: exploiting encoder representations for efficient
[24] Li X, Xia XG. A fast robust chinese remainder theorem-based phase unwrapping semantic segmentation. In: Proceedings of the IEEE visual communications and im-
algorithm. IEEE Signal Process Lett 2008;15:665–8. age (VCIP). St. Petersburg, FL, USA; 2017. p. 10–13.
11